AI Fleet Learning

One reason is the multiplicity of sensors they have, looking in every direction with optics and radar, plus they don't get angry, and they obey all traffic laws to the letter. But the biggest reason is “fleet learning.” Whenever anything does go wrong, engineers at headquarters examine all the factors and then modify the programming to mitigate that situation should it arise again - not just for the car in question, but for the whole fleet!  Uber drivers learn and get better with experience, but only the individuals, not the whole collection of Uber drivers.  Presumably, the Waymo learnings stay within the Waymo fleet, but they could be passed on to other autonomous taxi companies.

Let's apply this analogy to another profession - say, radiology. The process of learning radiology involves many years of college followed by more years of internship and clinical practice to achieve a minimum level of competence. To stay competent, radiologists must continue to read journal papers, go to conferences, read books, and practice other very slow and hit-or-miss methods of learning.  On the other hand, if AI radiologists were built into, say, MRI machines, and all of these AI entities were connected at some global hub, then any new knowledge or discovery could immediately be programmed or accessed by all the other AI-driven MRI machines. I am confident that this scenario will happen, if it hasn't started already.

In San Francisco, there will be competition among fleets of auto-taxis. Those fleets that learn the fastest and most fully will be the most profitable, and the riders will benefit from the increasing safety and lowered costs due to the competition.  In other professions, there will certainly be similar kinds of AI-driven “fleet competition” that will be very beneficial for humans, if they can afford it when they have no jobs.